SDRL for Hourly Electricity Load Forecasting: A Prophet–Gradient-Boosting Hybrid with Statistically Validated Accuracy (Published)
Accurate short-term load forecasting underpins unit commitment, economic dispatch, and electricity trading, and forecast error carries direct economic cost. This study proposes Sequential Decomposition–Residual Learning (SDRL), a three-stage hybrid in which a Prophet model extracts trend, multi-scale seasonality, and holiday effects; gradient-boosted trees model the remaining non-linear residual from 55 engineered features; and base learners are combined by an unweighted average. Evaluated on 145,392 hourly observations from the PJME region (2002–2018) under a strict temporal split reserving 40,201 hours for testing, the principal ensemble attains a mean absolute error of 182.92 MW (MAPE 0.563%), the lowest among thirteen benchmarked models, including two contemporary deep architectures. Its 15.52 MW advantage over the strongest non-decomposition baseline is significant under the Diebold–Mariano test (p < 0.001) and a moving-block bootstrap. A CPU-only variant attains 190.03 MW, and SHAP attribution elucidates the underlying mechanism, yielding an accurate, interpretable, and reproducible forecaster.
Keywords: Prophet, electricity load forecasting, ensemble forecasting, gradient boosting, interpretable machine learning, time-series decomposition